Skip to main content

Cloudflare Vectorize

If you're deploying your project in a Cloudflare worker, you can use Cloudflare Vectorize with LangChain.js. It's a powerful and convenient option that's built directly into Cloudflare.

Setup

Compatibility

Cloudflare Vectorize is currently in open beta, and requires a Cloudflare account on a paid plan to use.

After setting up your project, create an index by running the following Wrangler command:

$ npx wrangler vectorize create <index_name> --preset @cf/baai/bge-small-en-v1.5

You can see a full list of options for the vectorize command in the official documentation.

You'll then need to update your wrangler.toml file to include an entry for [[vectorize]]:

[[vectorize]]
binding = "VECTORIZE_INDEX"
index_name = "<index_name>"

Finally, you'll need to install the LangChain Cloudflare integration package:

npm install @langchain/cloudflare

Usage

Below is an example worker that adds documents to a vectorstore, queries it, or clears it depending on the path used. It also uses Cloudflare Workers AI Embeddings.

note

If running locally, be sure to run wrangler as npx wrangler dev --remote!

name = "langchain-test"
main = "worker.ts"
compatibility_date = "2024-01-10"

[[vectorize]]
binding = "VECTORIZE_INDEX"
index_name = "langchain-test"

[ai]
binding = "AI"
// @ts-nocheck

import type {
VectorizeIndex,
Fetcher,
Request,
} from "@cloudflare/workers-types";

import {
CloudflareVectorizeStore,
CloudflareWorkersAIEmbeddings,
} from "@langchain/cloudflare";

export interface Env {
VECTORIZE_INDEX: VectorizeIndex;
AI: Fetcher;
}

export default {
async fetch(request: Request, env: Env) {
const { pathname } = new URL(request.url);
const embeddings = new CloudflareWorkersAIEmbeddings({
binding: env.AI,
modelName: "@cf/baai/bge-small-en-v1.5",
});
const store = new CloudflareVectorizeStore(embeddings, {
index: env.VECTORIZE_INDEX,
});
if (pathname === "/") {
const results = await store.similaritySearch("hello", 5);
return Response.json(results);
} else if (pathname === "/load") {
// Upsertion by id is supported
await store.addDocuments(
[
{
pageContent: "hello",
metadata: {},
},
{
pageContent: "world",
metadata: {},
},
{
pageContent: "hi",
metadata: {},
},
],
{ ids: ["id1", "id2", "id3"] }
);

return Response.json({ success: true });
} else if (pathname === "/clear") {
await store.delete({ ids: ["id1", "id2", "id3"] });
return Response.json({ success: true });
}

return Response.json({ error: "Not Found" }, { status: 404 });
},
};

API Reference:

You can also pass a filter parameter to filter by previously loaded metadata. See the official documentation for information on the required format.